doi: 10.1093/comnet/cny007
A functional complexity framework for the analysis of telecommunication networks
Merim Dzaferagic†, Nicholas Kaminski, Neal McBride, Irene Macaluso and Nicola Marchetti
CONNECT, Research Centre, Trinity College Dublin, 2 Dublin, Ireland †Corresponding author. Email: [email protected]
Edited by: Piet Van Mieghem
[Received on 18 September 2017; editorial decision on 9 March 2018; accepted on 13 March 2018]
The rapid evolution of network services demands new paradigms for studying and designing networks. We propose a framework to investigate the underlying mechanisms of wireless network functions. This frame-work isolates and analyses a netframe-work function as a complex system. We propose functional topologies to visualize and systematically study the relationships between system entities. We also define a complexity metricCF(functional complexity), which quantifies the variety of structural patterns and roles of nodes in the topology. This complexity metric provides a wholly new approach to study the operation of telecommu-nication networks. We study the relationship betweenCFand graph structures by analysing graph theory metrics in order to recognize complex organizations.CFis equal to zero for both a full mesh topology and a disconnected topology. We show that complexity is high for a structure with shorter average path length and higher average clustering coefficient. We make a connection between functional complexity, robust-ness and response to changes that may appear in the system configuration. We also make a connection between the implementation and the outcome of a network function, which correlates the characteristics of the outcome with the complex relationships that underpin the functional structure.
Keywords: complex system science; functional complexity framework; telecommunication networks; structural complexity; frequency allocation.
1. Introduction
The transition of humanity into the information age has precipitated the need of new paradigms to comprehend and overcome a new set of challenges. Specifically, the telecommunications networks that underpin modern societies represent some of the largest scale construction and deployment efforts ever attempted by humanity, with renovations occurring nearly continuously over the course of decades. The result is networks that consist of numerous subsections, each of which following its own trajectory of development, commingled into a complex cacophony. Considering the high degree of heterogeneity and dense interplay of network elements in proposed 5G and internet of things (IoT) systems, achieving holistic understanding of network operation is poised to become an even more challenging prospect in the near future. The focus of our article is to provide the paradigms necessary for comprehension of the multi-faceted, intermingled and vital foundation of new networks by introducing a metric that quantifies the organizational structure of network functions.
understanding the function and thus the network itself. In order to analyse functional aspects of a telecom-munication network, we introduce a framework to map the network function into a functional topology. The functional topology enables a complex systems approach to analyse functions of telecommunications networks. We also introduce a metric which quantifies the complexity1of a particular implementation of
a network function. Specifically, the proposed metric analyses the underlying communication between network entities, which provide the network function.
The main contributions of this article are:
• We introduce a framework, which enables the functional analysis of a telecommunication network;
• We provide several examples that show how to map a network function into a functional topology;
• We provide a new complexity metric that quantifies the organizational structure of a telecommuni-cation network function.
The rest of the article is organized as follows. Section2presents the literature related to our work. In Section3, we introduce our network function framework, and we map two network functions into func-tional topologies. Section4introduces our complexity metric (i.e., functional complexity). In Section5, we examine the main properties of the functional complexity metric, and we investigate common graph structures that describe complex and non-complex functional relations. Section6concludes our analysis, and we additionally discuss current and future work related to this article.
2. Literature review
Using complex systems science, we can analyse network functions holistically. This interdisciplinary field draws attention from researchers in physics, mathematics, engineering and many others. As we plan future networks, the experience of these fields is useful to draw upon. Articles like [1–7] testify about the interdisciplinary nature of the complex systems analysis. The authors of [8] discuss issues of different research areas. This discussion leads to the conclusion that different scientific fields face similar problems. An interdisciplinary approach, which implies borrowing solutions from other scientific fields, would save time and effort that researchers put into solving problems.
Different sciences faced the problem of increasing complexity differently. In [9] the author presents a categorization of different complexity metrics and emphasises the interdisciplinary applicability of these metrics. The authors of [10–14] focus on community detection, which allows us to analyse the structural organization of a complex system. In [10] the authors investigate the underlying interactions between mobile phone users, which determine the affiliation to a community. Similarly, in [13,14] the authors focus on the modularity of the system. The authors of [11,12] focus on links rather than nodes, which enables the detection of overlapping communities. Additionally, their approach allows them to analyse the hierarchical structure of a complex system. In contrast, we focus on the structural features of the graph representation of a network function implementation, rather than on community detection.
The authors of [15–17] emphasize the importance of the evolution of tools, we use to analyse and understand complex systems. The authors of [15] and [16] focus on social networks and user behaviour. In [15], the authors propose a data-to-model process which allows them to analyse complex social interactions. This approach enables the prediction of the developments of eventual disasters in the system,
1 With the term ‘complexity’ we refer to a specific set of complex systems science quantities, related to the interactions between
which makes it possible to prepare the disaster recovery scenarios. The authors of [16] propose an agent-based framework for modelling competitive and cooperative behaviour under conflict [i.e., common-pool resource (CPR) Dilemma]. The framework allows them to study how and why we reach some outcomes, and to determine the conditions needed to achieve desirable outcomes in a complex system. In [17], the authors analyse the human travel patterns based on the trajectory of 100,000 mobile phone users. The understanding of the mobility patterns allows them to predict the movement and therefore the influence on spreading viruses, urban planning, mobile network planning, etc. We propose a framework that treats the implementation of a network function as a complex system, which allows us to focus on the network function itself, rather than the impact of the unpredictability of human behaviour on the network.
Due to our interest in telecommunication networks, articles like [18] and [19] are of particular interest to us because they emphasize the growing complexity of the next generation of communication networks, which demands the development of new frameworks and tools to model and analyse these networks. The authors of [18] propose an agent-based modelling framework to model and analyse the performance of sensor networks. While the authors of [18] focus more on the theory behind complex systems and agent-based modelling as a tool to model and analyse such systems, the authors of [19] apply agent-based modelling to a real world IoT problem. In [19], the authors analyse the impact of MAC protocol selection on communication performance in terms of spectrum utilization and accuracy of information.
Another approach to analyse and understand complex networks is graph theory. In [20–22], the authors emphasize the importance of graph theoretical tools to get a better understanding of the increasing com-plexity of systems. In [20], the author highlights the main difference between complex and non-complex graphs and introduces graph theoretical metrics to categorize, analyse and understand certain properties of these graphs. The author of [21] discusses how eigenvalues and structures of graphs are interrelated. The authors of [22], discuss different tools to model and extract knowledge from complex graph struc-tures. The main focus in [22] is graph mining, but the tools and metrics (e.g., graph generation, graph patterns, community structure, clustering, path-length) used to model complex relationships between entities are very important in complex systems science, and therefore, we compare our metric to some of these metrics.
Authors of [23–27] analyse telecommunications networks as complex systems. The approach in [23,24] involves analysis of complex phenomena in telecommunication networks (spreading patterns of mobile viruses and connection strengths between nodes in a social network), which are the result of complex user behaviour. The authors of [26,27] analyse the complexity of outcomes of a self-organizing frequency allocation algorithm. In [27], they analysed a relationship between robustness and complexity of the outcome. In contrast to the approach in [23–27] the work, we propose here targets the network function itself, which means that we analyse the mechanisms that enable complex user behaviour and system outcome. We use the same frequency allocation algorithm proposed in [27] to show an example of how to map a network function (i.e., frequency allocation) into a functional topology.
3. Network function framework
In order to specify and analyse the complexity of a network function, we introduce thefunctional framework. Our framework represents an abstraction of a telecommunication network modelling its operation by capturing all elements, i.e., nodes and connections, necessary to perform a given function. Our framework includesfunctional topologies, which are graphs created based on the functional connectivity between system entities. In contrast to logical topologies, which refer to data flows between network nodes, functional topologies depict the connectivity pattern of a network function, which implies a broader meaning to nodes and links.
A node in our topology represents a functional entity of a network node or any information source that is part of the given network function. The links indicate dependencies between nodes. The topology as a whole depicts the specific implementation of the network function. The topology allows us to apply our model to analyse the complexity of the implementation. The functional complexity quantifies the organizational structure by analysing the variety of relationships between system entities and roles that these entities have in the topology.
As an instructive example, we focus on self-organization from a frequency allocation perspective. More precisely, we use the frequency allocation algorithm from [27] to describe the process of mapping a network function into a topology and in the end to calculate the functional complexity.
The self-organizing frequency allocation algorithm considers a cellular network as a two-dimensional cellular automaton. Each cell in the model represents a self-organizing wireless system. The algorithm works based on the local information that nodes gather from their neighbours. Briefly, every cell senses the given frequency channels, and allocates a channel with no interference. For more details about the algorithm the reader is referred to [27].
We focus on two frequency allocation algorithms (self-organizing algorithm and random frequency assignment) from [27]. Herein, we analyse the implementation rather than the outcome of the function. Each autonomous network is modelled as a node in a lattice which means that the physical topology is the same for both implementations. The physical topology according to the Moore neighbourhood is shown in Fig.1. Different implementations of the network function are mapped into different functional topologies.
By analysing the physical topology shown in Fig.1, we recognize a motif that represents the Moore neighbourhood. The motif consists of nine neighbouring cells, as shown in Fig.2a. The motif depicts the local connections between neighbouring cells. The entire network is simply a repetition of that motif. The functional topology of an implementation of a network function is built upon the motif (Fig.2).
We start with the random frequency allocation algorithm. It assumes that every cell assigns the frequency completely randomly from the set of available frequency channels. In order to create the functional topology let us imagine a virtual decision maker entity that is moving from one cell to another. At every cell the decision maker entity has no information about the allocated frequencies of other cells, which means that there are no functional connections between any two nodes. The result is a functional topology represented with a non-connected graph (Fig.2b).
Fig. 1. The physical topology according to the Moore neighbourhood from cellular automata.
(a) (b)
Fig. 2. (a) The functional topology that represents the distributed self-organizing frequency allocation algorithm; the Moore neighbourhood motif. (b) The functional topology that represents the random frequency allocation approach. Independent network functions have no interconnections.
between them. By restricting the reachability to a certain number of hops along the topology we may examine the scale size of interest.
In [2] the authors discussed the relation between the scale size and the complexity, and they emphasize the importance of the scale size noting that the same system can show high complexity on one scale and low complexity on the other. The functional connectivity of each node represents the lowest scale in the functional framework and describes the interactions in the process of making a decision on this particular node. From the functional point of view the lowest scale size models a decision that is made by one node (in our case the frequency allocation of one cell). In order to apply a multi-scale approach we include higher scales in our analysis as follows. As every node represents a functional part, a subtask or an informational source, every group of nodes represents a group of subtasks that are part of the same function. The two-hop scale size considers the node, its neighbouring cells and neighbours of its neighbours. On this scale size the topology includes the interactions of nodes that are inside the radius of two hops. In our examples this scale size represents the frequency allocation of a two hop neighbourhood and the interactions between nodes in the process of making a decision among a group of nodes. Note that the reachability of nodes increases with the scale size. In this case every node can reach any other node that is two or less than two hops away. The highest scale size implies a full mesh topology, where every node can reach any other node in the topology. This approach enables a multi-scale functional analysis, which is presented in the results section.
4. Complexity model
The traditional reductionist approach attempts to explain an entire system in terms of its individual components. In contrast, complex systems analysis is based on the relations between system parts which result in a greater outcome than would be expected from a simple sum of the outcomes of the individual parts. Therefore, complex systems move the focus from the node to the network.
As we represent network functions with topologies, the complex systems approach allows us to analyse the relations between functional parts. Such analysis captures the joint effort of functionally interconnected system parts and provides a measurement of the deviation of the complex behaviour compared to a linear (non-complex) system. In order to capture this deviation, we analyse the joint effort of system sub-parts, which are represented with subgraphs of the functional topology. Different subgraphs with the same size capture the variety of organizational structures in the topology, whereas different subgraph sizes allow us to understand the gain from increasing number of nodes and interactions. In order to capture the non-linear joint effort of functional parts, we analyse the interactions of nodes involved in the decision making process on different scales. The scale sizerdetermines the maximum number of allowed hops between two nodes in order to reach each other. In modelling a network function, our focus from the complexity point of view is the degree of interconnectivity, which is represented by the reachability of functional parts.Ris the maximum scale size of the system, which is defined as the longest shortest path in the entire topology (i.e., the graph diameter of the functional topology). Ifr=R the reachability of every node in the topology is equal to 1.
In order to capture the relationships that underpin the operation of a given network function, we consider the interactions enabled by the structure of the functional topology on various scales of operation. Specifically, we consider that a node interacts if another node can reach this node on a particular scale. As such, we employ the Bernoulli random variablexnto describe the potential for a node to interact on a particular scale, under the assumption that interactions are uniformly initiated by all nodes within the topology. Therefore, the probability distribution ofxnis determined by the reachability of nodenfor a given scale sizer. Equation (4.1) provides the relative reachability of noden, wherein
Table 1 The notation used in the equations
Symbols Meanings
n Node
N Total number of nodes in the functional topology
j Subgraph size—number of nodes in the subgraph
r Scale size
R Maximum scale size, which is defined as the longest shortest path in the whole functional topology
j
k One of theksubgraphs withjnodes that is induced from the functional topology graph
in
r Number of nodes that can reach nodenfor a given subgraph
xn Bernoulli random variable.xn = 1 indicates that an interaction in the course of function operation involves noden, whereas xn = 0 indicates that the interaction does not involve noden, for a given scale
X Sincexnis a Bernoulli random variableX= {0, 1}
pr(xn=1),pr(xn=0) Probabilities that any given interaction in the course of function operation involves or does not involve nodenfor a given scale r
Hr(xn) Entropy of noden, which indicates the uncertainty of involvement of noden in the operation of a network function for a given scale sizer
of nodes that can reach noden for a given scalerandjrepresents the number of nodes for the given subgraph. Table1summarizes the notation used in our equations.
pr(xn=1)= in
r
j (4.1)
Shannon entropy for each node of the given subgraph with sizej, and scale sizeris calculated with equation (4.2).
Hr(xn)=
xn∈X
pr(xn)·log2 1 pr(xn)
(4.2)
The total amount of information of the kth subgraph with j nodes for scale r is calculated with equation (4.3).jkis one of theksubgraphs withjnodes. The total amount of information represents the total uncertainty, which is related to the actual roles of nodes that appear within a subgraph and different subgraph patterns.
Ir(jk)=
n∈jk
Hr(xn) (4.3)
The average amount of information for a given subgraph size jis calculated with equation (4.4).
βjin the equation represents the number of connected subgraphs with sizej.
Ir(j) = 1
βj
βj
k=1 Ir(
j
k) (4.4)
Functional complexity is calculated with equation (4.5).
CF= 1 R−1
R−1
r=1 N
j=1+r
Ir(j) −
r+1−j r+1−NIr(
N) (4.5)
Note thatNis the set of all nodes in the functional topology. Forj<1+rthe amount of information is equal to zero, because every subgraph with sizej<1+rfor the scale sizerrepresents a full mesh topology in terms of reachability, therefore,Hr(xn)=0 andIr(
j k)=0.
As shown in equation (4.5), in order to determine the complexity of an implementation of a net-work function, we compare this particular implementation(Ir(j))to a non-complex model of itself
(r+1−j r+1−NIr(
N)). A non-complex model is expressed by describing the functional parts, which provides
full information about such a network function because the function is the sum of the functionalities of its parts. Therefore, the non-complex model assumes that every functional part always contributes the same amount of information, meaning that the increasing number of analysed nodes leads to a linear increase in the total amount of information and in equation (4.5), it is represented by rr++11−−NjIr(N). Con-versely, a complex implementation implies that the network function relies on communication between its parts which results in a higher utility outcome, i.e., greater than the sum of its parts. Our functional complexity metric captures this difference between the amount of information for the complex and non-complex implementation, which is modelled with the inner sum in equation (4.5). Since we express the potential of a node to interact with other nodes as its reachability in the functional topology, Shannon entropy calculated based on this reachability represents the uncertainty of interactions for a subgraph of the functional topology. In equation (4.5), we capture the average uncertainty of interactions for all subgraph sizes (Ir(j)), and compare it to the uncertainty of interactions which is expected from the non-complex model, which is the functional complexity and in equation (4.5) it is represented by the inner sum. This means that the functional complexity captures how much more information exits in the interactions between nodes than it would be expected from a naive linear representation of these interac-tions. In engineering, non-linear problems are often approximated as linear problems for small changes in the input of the system. This approximation often results in models that have limited applicability, and the functional complexity quantifies how much information is lost by this linear approximation.
ofHr(xn). This again, leads to high values of Ir( j
k), which result in high functional complexity. The functional complexity is zero for a fully connected and for a disconnected graph. The reachability of each node in a fully connected and disconnected graph are equal to 1 and 0, respectively, which results in zero Shannon entropy and zero complexity.
The variety of roles emphasises the importance of interconnectivity of functional parts. Since nodes represent functional parts, different roles refer to different interactions that underpin the operation of a network function. If one node is highly connected to a group of nodes, this means that it influences the entire group. At the same time the group has a high influence on the operation of this node. In our frequency allocation example, the disconnected topology shows that every node is independent, which results in one simple role (disconnected node) for every subgraph. Describing the relationship between a single element and the rest of the functional topology provides all necessary information needed to understand the relationships of the entire topology. Therefore, this structure is not considered as complex. Conversely, the self-organizing algorithm provides different roles for nodes. This results in a complex structure, where the outcome depends on the communication between the structural parts, which means that the communication between system entities becomes more important to understand than the entities themselves.
5. Analysis
In this section, we apply our complexity model to several functional topologies to examine the main properties of the functional complexity metric. In addition to the two implementations of the frequency allocation function, we also investigate common graph structures (e.g., ring, star, bus, full mesh) that describe complex and non-complex functional relations. In the course of this analysis, we present the relationship between the complexity of a function and graph theory metrics that describe the functional topology representation of the network function.
A complex system, and hence a complex function, is determined by many parameters. We compare our metric to the following graph theoretic notions to highlight the utility of our methodology:
• Number of nodes,
• Average path length,
• Clustering coefficient.
Intuitively, if the number of nodes increases the system becomes more complex. The average path length indicates the distance between elements, which implicitly refers to the relationship strengths in the network. The clustering coefficient provides information about the overall strengths of functional dependencies between nodes grouped around a central entity. All these metrics provide useful information about certain organizational characteristics of the system. We start the analysis with common graph structures like a bus, ring, star and full mesh topology to investigate the impact of different graph structures on the amount of complexity. Figure3shows complexity values of theses topologies in the range of six to ten nodes. Additionally, we present the correlation between the complexity metric and different graph theory metrics.
5.1 Bus, ring, star and full mesh topologies
Fig. 3. Complexity of a bus, ring and star topology.
represented by a full mesh topology are non-complex. The structures of any subgraph in a full mesh topology are also full mesh connectivity patterns. This means that a full mesh topology represents a function in which every functional part interacts with all other entities in order to make a decision. The functional structures and roles of nodes in a graph that represents such functions are always the same (each individual node is a hub). This means that in order to describe the system (the function), it is enough to describe an individual element and such functions are considered as non-complex.
All the nodes in a bus topology are fairly spread out (i.e., the average path length between nodes is high). In other words, the bus topology represents a connectivity pattern with weak overall connections among parts of the topology. The bus topology does not provide a great variety of organizational structures, which results in low complexity. According to Fig.3, functional complexity of the bus topology increases with the increase in the number of nodes.
Figure3shows that the functional complexity of a ring topology is higher than the complexity of a bus topology for the same number of nodes. The ring topology conforms to the same trend in which complexity increases with the number of nodes. The ring topology compared to the bus is less spread out (i.e., the average path length is shorter) and therefore provides tighter connections between the nodes. As the nodes are closer to each other, due to the additional path, this functional topology represents tighter functional relationships compared with the bus topology. Tighter functional relationships result in higher complexity.
The star topology is a functional relationship in which one node interacts with all other nodes in the decision making process. It also conforms to the same trend in which complexity increases with the number of nodes in the topology. The star structure has tight connections between its parts. This means that the star topology depicts stronger functional relationships compared with the bus and ring topology. Therefore, the star topology has the highest complexity value compared with the other structures presented in Fig.3.
analyse the complex relationships between system entities. According to Fig.3, the star topology is the most complex structure compared with the bus, ring and full mesh topologies. Considering the sparse connectivity and short average path length between entities represented with a star topology, we believe that this type of relationship contributes to the increase in complexity. Based on the analysis of different sizes of ring and bus topologies, we believe that spreading (distributing) the functionality all over the topology also affects the increasing complexity values. We expect that a combination of these two types of relationships results in highly complex organizational structures. Such a structure consists of local centres (star subgraphs/clusters) that are highly interconnected. As functional complexity captures the variety of structural patterns, we expect that complexity is high if the number of local centres is high.
5.2 Comparison with graph theory metrics
In this subsection, we analyse the relationship between complexity and different graph theory metrics. We also want to emphasize that the graph theory metrics do not include the quantity that is measured by equation (4.5). In order to do so, we investigate the correlation between these metrics, and show that the graph theory metrics focus on individual aspects (e.g., clustering, connectivity, degree distribution) of the topology, whereas functional complexity quantifies the overall uncertainty of interactions between functional entities which emerge from the local interactions between them.
To describe the relationship between the average path length and the functional complexity of an implementation of a network function we analyse all possible graph organizations (distributions of links) between six nodes (see Fig.4). The Pearson product-moment correlation coefficient for these two metrics is−0.43 (see Table 2). Notice in Fig.4 that the highest functional complexity for six nodes is 2.9. Compared with the bus, ring and star topology the highest functional topology is at least 1.5 times higher. An important fact is that for the same average path length we have a range of different complexity values. Therefore, despite the high correlation between the average path length and complexity, we can’t predict
Table 2 Correlation between graph theory metrics and the functional complexity metric for all possible distributions of links between six nodes
Correlation variables Correlation
Average path length—complexity −0.43
Clustering coefficient—complexity 0.15
Avgerage path length–clustering coefficient—complexity 0.47
Fig. 5. The relationship between the clustering coefficient and functional complexity for every possible functional topology with six labelled nodes (we consider connected graphs only). We consider all unique combinations of edges starting with the minimum of five up to the maximum of fifteen. The Pearson product-moment correlation coefficient for these two variables (clustering coefficient and functional complexity) is 0.15.
the complexity of a structure based on the average path length. Additionally, the maximum value of complexity occurs for an extremely low average path length, which is still higher than one. Figure4 also shows that the envelope of the complexity values is a non-monotonic function. That is, there exist complexity values for longer average path lengths that are larger than all complexity values for some shorter average path lengths. Notice also that complexity is equal to zero for the average path length equal to one. This value of the average path length represents a full mesh topology.
Fig. 6. The relationship between the clustering coefficient, average path length and the functional complexity for every possible functional topology with six labelled nodes (we consider connected graphs only). We consider all unique combinations of edges starting with the minimum of five up to the maximum of fifteen. The complexity value is represented by the colour and diameter of circles.
that all neighbours of every node in the functional topology are neighbours to each other, which represents a full mesh topology.
In order to analyse the joint relationship between complexity and both of the graph theory metrics presented above (the average path length and the average clustering coefficient) we plot all these metrics together on one graph (see Fig.6). Thexandyaxis represent the average clustering coefficient and the average path length, respectively and the complexity is represented by the radius of the circles. Addition-ally, the circles are coloured differently to emphasize the different complexity values. The multivariate correlation between these three variables is 0.47 (see Table2). The overlapping circles with different radii mean that for the same average path length and average clustering coefficient there exist multiple topologies with different functional complexity values. The functional complexity is an information the-ory metric, while the average path length and the average clustering coefficient are graph thethe-ory metrics, meaning that they quantify different properties, and therefore this was expected. As shown in the analysis above, graph theory metrics allow us to analyse individual aspects of the functional topology, whereas complexity adds additional information about the diversity of relationships between functional parts and provides a different approach to analyse complex network functions.
Fig. 7. The relationship between the correlation coefficients of the graph theory metrics (average path length, clustering coefficient) and the complexity metric for different sizes of functional topologies. These values were calculated for all unique combinations of
nnodes and fromn−1 to(n∗(n−1))/2 edges.
5.3 Complexity of the frequency allocation algorithms
In order to analyse the functional complexity of the frequency allocation algorithms presented in the framework section, we apply our complexity model to the functional topology representation of these network functions. Our complexity metric, when applied over the functional topology, gives insight into the functional relationships between system entities, which allows us to focus on the joint effect of functional parts in order to execute the modelled network function.
The random frequency allocation algorithm is represented by a disconnected graph. Each element represents a closed system which results in the absence of connections between them. From the system perspective the functional parts of this implementation represent a set of independent elements. As the elements do not interact with each other, in order to describe the set we can simply describe its elements. More precisely, to describe the set we have to describe only one element, because each element of the set is the same. Each element of the set simply assigns one of the available frequencies. Such a set that is represented with independent elements has zero functional complexity.
Finally, the self-organizing distributed frequency allocation algorithm is represented by the functional topology shown in Fig.2a. In order to analyse the complexity of this implementation of the frequency allocation function, we apply our complexity model to the functional topology representation. Figure8a depicts the deviation in the amount of information with increasing subgraph sizes from a linear increase (i.e., the functional complexity). In equation (4.5) the linear increase is represented by(rr++11−−NjIr(N)). As shown in equation (4.5), it is calculated as the uncertainty which is expected from the calculation performed on the whole system (Ir(N)). The linear increase represents an equivalent simple system implementation, meaning that all the subgraphs of a specific size feature the same uncertainty of inter-action. Every functional part contributes with the same amount of information, which results in a linear increase over increasing subgraph sizes.
Fig. 8. The functional topology has nine nodes; the maximum scale sizeRis 2; the functional complexity is represented by the area between the linear and non-linear function. (a) The relationship between the uncertainty of interactions for all subset sizes (Ir(j)) and the uncertainty, which is expected from the calculation performed on the whole system(rr++11−−NjIr(N)). (b) The distance between the uncertainty of interactions for all subset sizes and the uncertainty, which is expected from the calculation performed on the whole system(I(j) = |Ir(j) − rr++11−−NjIr(N)|), i.e., the functional complexity of the self-organizing frequency assignment approach.
subgraph size (Ir(j)), and it represents the uncertainty of interaction among nodes for this specific subgraph size. According to equation (4.5), complexity is calculated as the sum of differences between the average amount of information for a specific subgraph size and the amount of information which is expected from the calculation performed on the whole system (the non-linear and linear functions in Fig.8a).
Considering that for the functional topology of the self-organizing algorithmR=2, the multi-scale approach applied on this functional topology implies only a single scale analysis, that isr = 1. The subgraph sizes of interest are in the range of 2–9. Considering thatr=1, subgraphs with the size two or smaller represent full mesh topologies. The maximum value of the subgraph size is nine because the functional topology consists of nine nodes. The complexity of the self-organizing frequency allocation algorithm is 1.69 and it is represented by the area between the linear and non-linear functions in Fig.8a. Figure8b shows the distance between the uncertainty of interactions for all subset sizes and the uncertainty which is expected from the calculation performed on the whole system, i.e., the functional complexity of the self-organizing frequency allocation approach. More precisely, Fig.8b depicts the single scale complexity function, which is represented by the inner sum of equation (4.5). The functional complexity is the sum of the function depicted in Fig.8b.
than the output of the network function. As discussed above, the functional complexity of the random frequency allocation algorithm is zero, and the functional complexity of the self-organizing distributed frequency allocation algorithm is 1.69. This suggests that a complex implementation of a function results in a complex outcome. In [27] the authors show that the response to a change of a system with highly complex output happens in a more manageable fashion causing less disruption. The authors of [27] also emphasize the higher robustness of a complex outcome compared to a non-complex outcome. Making a connection between the implementation and the outcome of a function correlates the characteristics of the outcome with the complex relationships that underpin the functional structure.
6. Conclusion
The growing size and heterogeneity of telecommunication networks leads to a need to change the way we analyse and model them. Also the rapid evolution of network services demands a new approach to analyse them. The aim of this article was to contribute to this new way of studying networks. We focus on network functions as building blocks of services. We consider network functions as complex systems. In order to provide a new approach to analyse and understand the impact of complex functional relationships between system entities, we developed a new framework. The framework allows us to visualise an implementation of a network function with graphs called functional topologies. We provide several examples that show how to map a network function into a functional topology. The graph visualisation of the implementation allows us to focus on the relationships between entities rather than the entities themselves.
The next step after mapping an implementation of a network function into a functional topology was to provide a metric that quantifies the organizational structure of the topology. Our functional complexity (CF) captures the variety of roles that each node in the topology has and the variety of structural patterns present in the topology.CF quantifies the deviation of the implementation of a network function from the non-complex model of itself. The quantification of this deviation as presented in Section4provides a new approach to understand increasingly complex telecommunication networks.
In order to study the impact of different structural patterns on the functional complexity in Section5, we start by analysing the complexity of simple graph structures (bus, ring, star and mesh). Additionally, we provide a detailed study that investigates the impact of several graph theory metrics on the functional complexity. We investigated the correlation between the combination of the graph theory metrics and complexity, due to the absence of high correlation between any graph theory metrics and the functional complexity. This analysis allowed us to make conclusions about the organizational structures that is needed in order to achieve high complexity.
In this article, we also analysed the functional complexity of two different implementations of the frequency allocation function (random and self-organizing). First, in Section3we explain how to map these implementations into functional topologies, and in Section5, we apply our complexity metric to quantify the functional complexity of these implementations. We showed that the random frequency allocation has zero complexity (and represents therefore a non-complex implementation), whereas the self-organizing implementation shows higher complexity. This allows us to make a connection between our results and the results from [27], where the authors calculated the complexity of the outcome of these implementations and got the same kind of results. This implies that a complex implementation results in a complex outcome of the function and as a part of our future work, due to the similar trends exhibited by both complexity metrics, we are confident that we can make justified assumptions about the robustness and response to change similar to [27].
use the functional topology framework as a part of our current work on distributed optimization in cellular networks to model the relationships between eNodeBs. We are also working on sampling techniques for functional topologies which will allow us to apply the framework to large networks. As a part of our current research we are also applying the functional framework to model the self-organized time division multiple access algorithm in the very high frequency (VHF) maritime mobile band for universal shipborne automatic identification systems. Our end goal, after investigating the relationships between functional complexity and different network key performance indicators (KPIs), is to use functional complexity to tune the tradeoff between these network KPIs.
The main advantage of using a metric like functional complexity instead of KPIs to tune and optimize the network is that functional complexity can be computed offline without relying on data collected in the network, which speeds up the network optimization process. On the other hand, as shown in [32], the first step in the process of calculating a KPI is to collect measurements. In order to collect the measurements the network has to be operational, meaning that the process of optimizing certain aspects of the network is possible only by monitoring the real network dynamics. As shown in the article, the functional topologies capture the relations between functional parts of the network based on the rules that are defined by the communication protocols themselves. In other words, the functional topology captures certain network dynamics without relying on measurements. Therefore, the functional complexity can be calculated even if the network is offline. By showing high correlation between the complexity metric and network KPIs (e.g., [30,31]), we show that a network operator would be able to optimize the network aspects related to those KPIs before the network is operational.
Overall, our complexity metric quantifies telecommunication networks in terms of their functional relationships and provides a wholly new approach to understanding the operation of networks. More precisely, the complexity metric quantifies the relationships employed during operation of network func-tions. This provides a new approach to study networks, which is especially relevant for next generation networks.
Funding
This work was supported by the Science Foundation Ireland under (13/RC/2077).
References
1. Liu, Y.-Y., Slotine, J.-J. & Barab ´asi, A.-L.(2013) Observability of complex systems.Proc. Natl. Acad. Sci. USA,110, 2460–2465.
2. Bar-Yam, Y.(2004) Multiscale complexity/entropy.Adv. Complex Syst.,7, 47–63.
3. Yaeger, L.(2007) Evolution Selects for and Against Complexity. Available at: http://vw.slis.indiana.edu/talks-fall07/Yeager.pdf.
4. Tononi, G., Sporns, O. & Edelman, G. M. (1994) A measure for brain complexity: relating functional segregation and integration in the nervous system.Proc. Natl. Acad. Sci. USA,91, 5033–5037.
5. Lopez-Ruiz, R., Mancini, H. L. & Calbet, X.(1995) A statistical measure of complexity.Phys. Lett. A,209, 321–326.
6. Balduzzi, D. & Tononi, G. (2008) Integrated information in discrete dynamical systems: motivation and theoretical framework.PLoS Comput. Biol.,4.
7. Joshi, N. J., Tononi, G. & Koch, C.(2013) The minimal complexity of adapting agents increases with fitness. PLoS Comput. Biol.,9, e1003111.
8. Evans, T. P., Ostrom, E. & Gibson, C.(2002) Scaling issues with social data in integrated assessment modeling. Integrated Assessment,3, 135–150.
10. Wang, Z., Zhang, D., Zhou, X., Yang, D., Yu, Z. & Yu, Z. (2014) Discovering and profiling overlapping communities in location-based social networks.IEEE Trans. Syst. Man Cybern. Syst.,44, 499–509.
11. Ahn, Y.-Y., Bagrow, J. P. & Lehmann, S.(2010) Link communities reveal multiscale complexity in networks. Nature,466, 761–764.
12. Evans, T. S. & Lambiotte, R.(2009) Line graphs, link partitions, and overlapping communities.Phys. Rev. E.,
80, 016105-1, 016105-8.
13. Newman, M. E. J. & Girvan, M.(2004) Finding and evaluating community structure in networks.Phys. Rev. E,
69, 026113.
14. Newman, M. E. J.(2006) Modularity and community structure in networks.Proc. Natl. Acad. Sci. USA,103, 8577–8582.
15. Lanham, M. J., Morgan, G. P. & Carley, K. M.(2013) Social network modeling and agent-based simulation in support of Crisis De-escalation.IEEE Trans. Syst. Man Cybern. Syst.,44, 103–110.
16. Bristow, M., Fang, L. & Hipel, K. W.(2014) Agent-based modeling of competitive and cooperative behavior under conflict.IEEE Trans. Syst. Man Cybern. Syst.,44, 834–850.
17. Gonz ´alez, M., Hidalgo, C. & Barab ´asi, A. L.(2008) Understanding individual human mobility patterns. Nature,453, 779–782.
18. Batool, K. & Niazi, M. A.(2017) Modeling the internet of things: a hybrid modeling approach using complex networks and agent-based models.Complex Adapt. Syst. Model.,5, 4–23.
19. Kaminski, N. J., Murphy, M. & Marchetti, N.(2016) Agent-based modeling of an IoT network.IEEE Int. Symp. Syst. Eng., pp. 1–7.
20. Newman, M.(2003) The structure and function of complex networks.SIAM review,45, 167–256.
21. Butler, S.(2008) Eigenvalues and structures of graphs.Ph.D. dissertation.
22. Chakrabarti, D. & Faloutsosc, C.(2008)Graph Mining: Laws, Generators and Tools. Lecture Notes in Computer Science: Springer-Verlag.
23. Wang, P., Gonzalez, M. C., Hidalgo, C. A. & Barabasi, A.-L.(2009) Understanding the spreading patterns of mobile phone viruses.Science,324, 1071–1076.
24. Onnela, J.-P., Saram ¨aki, J., Hyv ¨onen, J., Szab ´o, G., Lazer, D., Kaski, K., Kert ´esz, J. & Barab ´asi, A L.(2007) Structure and tie strengths in mobile communication networks.Proc. Natl. Acad. Sci. USA,104, 7332–7336.
25. Beigy, H. & Meybodi, M. R.(2010) Cellular learning automata with multiple learning automata in each cell and its applications.IEEE Trans. Syst. Man Cybern. Syst.,40, 54–65.
26. Macaluso, I., Cornean, H., Marchetti, N. & Doyle, L.(2014) Complex communication systems achieving interference-free frequency allocation.IEEE Int. Conf. Commun., 1447–1452.
27. Macaluso, I., Galiotto, C., Marchetti, N. & Doyle, L.(2016) A complex systems science perspective on wireless networks.J. Syst. Sci. Complex.,29, 1034–1056.
28. Kaminski, N. J.(2014) Social Intelligence for Cognitive Radios. Ph.D. Dissertation, Virginia Polytechnic Institute and State University, Blacksburg, Virginia.
29. Newman, M.(2010)Networks: An Introduction.New York, NY: Oxford University Press, Inc.
30. Dzaferagic, M., Kaminski, N. J., Macaluso, I. & Marchetti, N. (2017) Relation between functional complexity, scalability and energy efficiency in WSNs.Int. Wirel. Commun. Mobile Comput. Conf., pp. 675–680.
31. Dzaferagic, M., Kaminski, N. J., Macaluso, I. & Marchetti, N.(2017) How functional complexity affects the scalability-energy efficiency trade-off of HCC WSN clusteringIEEE Wirel. Commun. Network. Conf.